Architecture Catalog

Enerzyme supports several internal architectures and external wrappers. Choose based on targets (energy/forces only vs charge/dipole), system size, equivariance needs, and optional dependencies.

Internal architectures

Architecture	Charge	Dipole	Modular	Shallow ens.	Notes
SchNet	yes	yes	partial	yes	Good baseline
PhysNet	yes	yes	yes	yes	Electrostatics, D3/D4 optional
SpookyNet	yes	yes	yes	yes	Enzyme-scale clusters
MACE	yes	yes	partial	yes	Equivariant, higher cost
AlphaNet	varies	varies	partial	varies	See config TODOs

External wrappers

Architecture	Extra install
NequIP	nequip
XPaiNN	XequiNet and dependencies

External models are declared under Modelhub.external_FFs with the same active / layers pattern where supported.

Selection guidelines

Baseline / tutorial

SchNet — minimal dependencies, charge-aware stacks available.

Production QM-labeled clusters with charge and solvent

PhysNet or SpookyNet — long-range electrostatics, optional dispersion layers.

Maximum accuracy on diverse geometries

MACE or NequIP — equivariant message passing; tune cutoff and depth.

Active learning with force variance

Any architecture with ShallowEnsembleReduce or committee_size > 1.

Spin and charge

Charge-aware stacks need Q (and often ChargeConservation). SpookyNet-style models may use ElectronicEmbedding for charge and spin (S / multiplicity). Match simulation System.charge and multiplicity to training data conventions.

Reference configs

Full multi-architecture examples: enerzyme/config/train.yaml. Enable one FF entry at a time when starting (active: true).